Kimi Releases K3.1 Pro Flagship, Claiming Top Open-Source Rank on AI Index

The model, which appeared in an AI model arena under the alias Kibitz before launch, has 3 trillion parameters and supports 1 million tokens of context. Kimi said it plans to open the full model weights before Nov. 4. A Highspeed version with the same intelligence level but two to three times the output speed of the standard version was also released. K3.1 Pro is available on Kimi's website and App, Kimi Code, CLI and API.

On code, agent and knowledge benchmarks, K3.1 Pro trailed the strongest closed-source models but came close, according to the report. On Artificial Analysis' Intelligence Index, it rose to fourth globally and first among open-source models. Kimi also said the model outperformed GPT-6 Astra and Opus 5.5 on Office QA Pro v2 and SpreadsheetBench 2, two external benchmarks for financial document analysis and spreadsheet operations including financial modeling, cross-sheet formulas and error diagnosis.

In Ifanr's tests, K3.1 Pro generated 3D scenes from single prompts, including a Three.js and SVG car film showing a white Roadster on a C-shaped track, a rainy ancient city frozen by a sword strike and a monarch butterfly forest in Mexico's highlands. Kimi also launched 3D Make, which lets users preview and edit 3D models online and download files in formats such as blend and glb for 3D printing, or export them as video or live photos for social media.

The model also turned a room photo into a rotatable 3D webpage and generated a Blender-scripted robot asset for an interactive 3D product launch. Ifanr said K3.1 Pro handled tasks built around a specialist Skill, such as explaining how undersea cables connect the global internet, and produced a 15-second motion-graphics video from a one-sentence prompt. It also replicated Arc browser features including tab splitting, a sidebar, global search and spaces after inspecting an installation package, the report said.

The release comes as Chinese model makers have spent recent months emphasizing cost and speed, starting with DeepSeek V4 Flash, while closed-source frontier models such as Claude Opus 5.5 and GPT-6 Astra have continued to update. Kimi's K3, released in July, raised the scale of Chinese open-source models to 2.8 trillion parameters. Ifanr said K3.1 Pro has not comprehensively surpassed closed-source flagships but can now compete with them on several tasks, and it represents a continuing bet on larger models rather than only cheaper, smaller ones.

Ifanr said benchmark scores close to flagships do not always translate into equal performance in open-ended real-world tasks, where larger models may retain advantages in insight, imagination and language. It said real tasks such as large codebases and complex agent workflows require sustained planning, tool use and understanding of goals.